Multi-Dimensional Sparse Models.

Area of Science:

• Signal Processing
• Machine Learning
• Data Science

Background:

• Traditional sparse representation models convert multidimensional (MD) signals into 1D vectors.
• This 1D approach struggles with high-dimensional data, leading to inefficiencies in computational resources and memory usage.
• It also ignores the inherent tensor structure and diversity of MD signals.

Purpose of the Study:

• To propose novel MD synthesis/analysis sparse models for effective and efficient MD signal representation.
• To leverage the multilinearity of tensors for creating redundant bases under sparsity constraints.
• To develop corresponding dictionary learning algorithms and unified MD signal restoration formulations.

Main Methods:

• Utilizing tensor multilinearity to establish redundant bases for multi-linear maps with sparsity constraints.
• Developing MD sparse models that capture dimensional features through simultaneous and collaborative dictionaries.
• Proposing unified dictionary learning algorithms and signal restoration formulations for MD signals.

Main Results:

• MD sparse models outperform state-of-the-art 1D models in signal representation quality, computational overhead, and memory storage.
• Demonstrated effectiveness on MD signal denoising, image super-resolution, and texture classification tasks.
• The proposed MD models generalize 1D sparse models and adapt to signal properties.

Conclusions:

• MD sparse models offer a superior approach to representing high-dimensional signals compared to traditional 1D methods.
• The developed models and algorithms provide significant improvements in efficiency and representation quality.
• These MD models are flexible and applicable to various MD signal processing tasks.